The topline. San Juan, October 24, 2026. An AI triage tool reads the chart the way it is written, and a new study shows that the wording can matter as much as the vital signs. Researchers at UT Southwestern tested three open source language models on emergency department records and found that adding stigmatizing details to a patient’s history sometimes caused the models to lower the priority of patients who later deteriorated. The lesson is practical, not pessimistic: when clinicians write clean, neutral notes, they are also calibrating the tool that will help them sort the next waiting room.

Key facts

The team used patient records from emergency departments at Stanford Medicine and Beth Israel Deaconess Medical Center. They paired patients with similar initial severity, then asked three models to predict who was more likely to deteriorate and who should be prioritized. Before demographic, social, or potentially stigmatizing information was added, all three models identified higher risk patients at rates well above chance. The tools could do the job.

Then the researchers changed only the language. They compared neutral phrasing with stigmatizing phrasing for the same history: “established patient with multiple chronic conditions” against “17 visits in the past 12 months,” and “opioid use disorder, on maintenance therapy” against “active IV drug use, heroin, last use today.” Details associated with clinical stigma often shifted the models’ priorities, and in some instances nearly 10 percent of patients were incorrectly downgraded. The strongest effects followed two labels: frequent emergency department use and a history of psychiatric emergencies, mirroring a known human triage problem in which patients labeled as frequent visitors receive lower priority.

Clinical relevance

This study is best read as a specification for safe deployment, not as a reason to shelve triage tools. A second reader is only as good as the record it reads, and the chart is the clinician’s instrument twice over: it records the patient, and it now feeds the assistant. Writing “opioid use disorder, on maintenance therapy” instead of a stigmatizing equivalent is already good practice for the human reader. This study shows it is also good practice for the machine reader, because the model inherits the associations in the words it is given. The authors’ prescription follows directly. Bias varied by model, so no triage tool should be adopted on vendor claims alone; each one should be tested against real patient outcomes in the setting where it will run, with stigmatizing language variants included in the test set. The clinician stays the decision maker, and the tool earns its place by proving it helps that decision under realistic chart conditions, including the messy ones.

10%

That is the approximate share of patients incorrectly downgraded in some instances after stigmatizing information was added to their records. The figure is a worst case slice, not an average across the study, and it should be read that way. Even as a slice, it is the number that makes the abstract risk concrete: a single added detail sometimes shifted a model’s priority away from the patient who later deteriorated. First author Philip Jarrett’s summary is the takeaway for the charting clinician. Wording matters, because AI models process the words used to describe patients, and stigmatizing language can steer care away from the patients who need it most.

What we’re reading

  • npj Digital Medicine: the UT Southwestern study of language models in emergency triage.
  • UT Southwestern Medical Center: the study announcement with the paired phrasing examples.
  • Stanford Medicine and Beth Israel Deaconess Medical Center: the emergency department record sources behind the analysis.